Unet3+ uses full-scale skip connections to fuse feature maps of different scales, and learns feature representation from multi-scale aggregated feature maps through deep supervision. In addition, a hybrid loss function is proposed to combine classification tasks and segmentation tasks. It can enhance organ boundaries and reduce over-segmentation of non-organ images, resulting in more accurate segmentation results.
1. Full-scale Skip Connections
In order to make up for the inability of Unet and Unet++ to accurately segment the position and boundary of organs in the image, each decoder in Unet3+ combines the features of all encoders. These features of different scales can obtain fine-grained details and coarse-grained semantics.
2. Full-scale deep supervision
To learn hierarchical representations from full-scale aggregated feature maps, Unet3+ further employs full-scale deep supervision.
3. Loss function
4. Classification-guided module
In order to prevent over-segmentation of non-organ images and improve the segmentation accuracy of the model, an additional classification task is added to predict whether the input image has organs, so as to achieve more accurate segmentation.